ARCTIC: A knowledge distillation approach via attention-based relation matching and activation region constraint for RGB-to-Infrared videos action recognition

Zhenzhen Quan, Qingshan Chen, Yujun Li, Zhi Liu, Yan Cui

Abstract: Highlights•To enhance the precision of knowledge screening across different modalities, we construct attention-based relation matching modules within different feature layers of the teacher and student networks.•To reduce the discrepancy between different modalities, we consider the spatial activation region consistency constraint to keep the consistency between the most significant features of the teacher network and the student network.•To obtain more advantageous complementary information, we employ knowledge distillation loss to favor the selection of accurate predictions to enhance network performance.•We conduct extensive experiments on NTU RGB+D and PKU-MMD datasets, and the results show that our method is effective and feasible.